# what changes when intelligence becomes cheap ## innovation, marketing, entrepreneurship ![[cheap intelligence.png]] denys holovatyi, founder & ceo @ osnova gmbh https://osnova.services/ --- # contents - [[#innovation]] - [[#marketing]] - [[#entrepreneurship]] --- # artificial intelligence changes the cost of action ai makes it cheaper to: - test an idea - build a prototype - create and compare alternatives - automate a task - coordinate work - start a company but when execution becomes cheap, choosing the right thing becomes more important. --- # innovation ## from planning products to running experiments --- ## innovation becomes user-driven experimentation software used to be expensive enough that companies tried to understand everything before building. ai tools make it possible to build first versions quickly, put them in front of users, observe what happens, and change them again. the innovation loop becomes: 1. observe a real problem 2. build the smallest useful intervention 3. test it inside the user's real environment 4. measure behavior and quality 5. feed every response back into the product ![[Pasted image 20251117184857.png]] [[15 principles of ai product design]] --- ## automate what people hate the best discovery question is often not: > where can we use ai? but: > what task in your daily work do you hate? people have an emotional reaction to work they want to delegate. that gives us a problem, motivation, and an adoption hypothesis at the same time. [[15 principles of ai product design#automate what people hate]] --- ## start with familiar tools innovation fails when the user must adopt five new tools before receiving any value. the first ai product may therefore look like: - an email - an excel file - a form - an extension inside sap, jira, salesforce, or service now - a workflow connected to power bi or tableau the technology can be new while the interface remains familiar. ![[Pasted image 20251117002428.png]] [[15 principles of ai product design#start with familiar tools]] --- ## last-mile automation is the unexplored territory standard software automates the standard process. humans remain the glue between systems: - copy from a browser into a document - download, rename, unzip, and move files - drag an image into a presentation - check an exception in one system and approve it in another these tiny, irregular actions are too specific for conventional products and too changeable for rigid workflows. this is **last-mile automation**. ![[banana bruegel delivery banner.png]] [[last mile automation]] --- ## build simple things first many ai products can be tested before they are engineered: 1. one prompt with an attached document 2. the same prompt inside an email or backend automation 3. an integration that reads and writes to the source system 4. only then: dashboards, permissions, branching agents, and a full application rough prototypes invite feedback. every feedback item can become a prompt, a specification, a test, and a product change. ![[Pasted image 20251117003136.png]] [[15 principles of ai product design#build simple things first]] --- ## ai should learn from the process, not only the result language models see millions of code repositories. but a final repository is only the final answer. it hides: - the original requirement - failed attempts - reviews and corrections - changing constraints - the sequence in which the solution emerged commit histories contain this missing process. ![[Pasted image 20251117005138.png]] [[reverse engineering commit history to train reasoning models on process rewards]] --- ## reverse engineer a chain of commits for every commit: 1. take the repository at the previous state 2. inspect the diff, commit message, issue, pull request, and review 3. generate the instruction that could have produced the change 4. run the instruction against the previous state 5. compare the generated result with the real next commit the repository becomes an evolving sequence of states interleaved with prompts. not just a code dataset, but a dataset of how software is made. ![[Pasted image 20251117005347.png]] --- ## commit history can produce process rewards each generated step can be evaluated through: - the resulting diff - unit and integration tests - build success - static analysis - an llm-as-judge comparison - consistency with the actual next repository state this creates verifiers for intermediate work, not only a reward for the final answer. the larger innovation principle is transferable: > preserve the trace of how good work evolves. the trace may be more valuable than the final artifact. ![[your software factory – Horizontal.jpg]] --- # marketing ## infinite production meets finite attention --- ## the big promise of ai in marketing ai promises to automate almost the whole marketing pipeline: - research markets and customers - identify and score leads - personalize messages - generate campaigns and content - create images and video - run experiments - analyze performance - optimize the next campaign the promise is not simply faster copywriting. it is a closed learning loop from market signal to message to response. ![[02 - fire your marketing team.png]] --- ## most implementations stop at producing more 10 people pitch ai agents. almost everyone sells: 1. lead lists and automated outreach 2. endless marketing articles 3. workflows that look convincing in a demo but a sent message is not a qualified lead. an article is not attention. activity is not demand, trust, revenue, or retained business. ![[3 in 10 agents.png]] [[li - 03 - 10 people pitched me their AI agents]] --- ## people are tired of slop when every company uses the same models, the same prompts, and the same playbooks, brands begin to sound the same. the result is: - identical hooks - identical listicles - identical optimism - identical vocabulary - identical synthetic images - identical calls to action content volume grows while distinctiveness collapses. ![[2026-06-07_ai-parrot_cr.png]] [[i miss the time when ai was simply a parrot]] --- ## language models are central-tendency machines a language model learns a probability distribution over possible next tokens. under ordinary decoding, high-probability continuations are favored. that does **not** mean every output is a mathematical average. but without strong context, taste, constraints, examples, or deliberate exploration, the model gravitates toward patterns that are common in its training data. in marketing, the most probable message is often the least memorable one. --- ## average content is cheap — distinction is scarce ai is excellent at: - generating alternatives - adapting format and length - translating and localizing - extracting themes from research - testing combinations at scale humans and organizations must still contribute: - a point of view - proprietary evidence - taste - lived experience - a recognizable voice - the courage to exclude most possibilities ai can multiply distinction. it cannot manufacture distinction from an empty brief. --- ## marketing should begin with value, not content the useful test is: - is the pain obvious? - is the customer specific? - is the promise concrete? - is the return measurable? - can it be tested within weeks? the goal is not to replace a marketing team in three hours. the goal is to connect a real customer problem with a credible reason to choose you. [[li - 02 - fire your marketing team]] --- ## narrow distribution beats generic reach a focused go-to-market strategy can begin with: - one product environment - one community - one recognizable problem - a handful of trusted creators or partners - the first 1,000 users distribution is part of the product hypothesis, not a problem to solve after the product is finished. [[go to market in obsidian]] --- # entrepreneurship ## smaller teams, larger capability --- ## the minimum efficient company is shrinking as ai can complete longer tasks, fewer people are required to produce a given outcome. small teams can increasingly combine: - product development - software engineering - research - design - support - operations - marketing headcount becomes a weaker proxy for capability. revenue and useful output per employee become more important. ![[Pasted image 20251021230326.png]] [[future ai-enabled organizations]] --- ## early ai companies already show unusual leverage ![[Pasted image 20251021231621.jpg]] the snapshot in the vault estimates: - cursor: $3.3m revenue per employee - midjourney: $2m - openai: $1.5m - bolt: $1.4m - lovable: $526k - elevenlabs: $474k these figures are directional estimates from flashpoint and dealroom, not audited financial statements. the important signal is the organizational pattern: small teams can now reach meaningful revenue unusually quickly. --- ## new companies do not need an ai transformation large organizations must change: - legacy systems - procurement rules - permissions - incentives - roles - workflows - culture a new organization can start with ai embedded in every function. this is why one-person and very small companies can be viable in domains that previously required departments. ![[Pasted image 20251021232117.png]] [[future ai-enabled organizations]] --- ## the cost of intelligence is falling in march 2025, the original article observed a dramatic decline in the price of model inference and predicted another enormous reduction. ![[dane-vahey-of-openai-says-the-cost-per-million-tokens-has-v0-q_jUcKIIY2VaNBuDK5ppPcGQPUOrSE1zpXfG9QYjfsU.webp]] the exact **100,000x within two years** forecast was deliberately provocative and should not be treated as an established fact. the durable point is economic: > when the price of a useful capability falls, demand and the number of viable applications rise. [[cost of intelligence]] --- ## in july 2026, intelligence has a price ladder public api prices per 1 million tokens, 22 july 2026: | model | input | output | |---|---:|---:| | deepseek v4 flash | $0.14 | $0.28 | | gemini 3.5 flash-lite | $0.30 | $2.50 | | openai gpt-5.6 luna | $1.00 | $6.00 | | anthropic sonnet 5* | $2.00 | $10.00 | | openai gpt-5.6 sol | $5.00 | $30.00 | *anthropic introductory price through 31 august 2026; announced standard price is $3 / $15. sources: [deepseek](https://api-docs.deepseek.com/quick_start/pricing/), [google](https://ai.google.dev/gemini-api/docs/pricing), [openai](https://developers.openai.com/api/docs/models), [anthropic](https://claude.com/pricing) --- ## cheaper tokens do not make every business economical token price is only one component of the cost of an ai product. the real unit economics include: - repeated reasoning and agent loops - tool calls and search - data preparation and retrieval - evaluation and human review - failed runs and retries - latency and infrastructure - integration and maintenance - customer acquisition and support the economically correct question is not: > how cheap is the model? but: > how much verified customer value do we create per complete task? --- ## technology is not the value proposition banks survived multiple technological revolutions because their core business remains the supply and demand of capital. technology changes how quickly and efficiently the value is delivered. it does not automatically create the value. the same is true for ai companies: - an agent is not a value proposition - automation is not a business model - a model wrapper is not a defensible company [[AI will create 8 industries and destroy some more]] --- ## ai-native entrepreneurship begins with four questions 1. what painful outcome does the customer need? 2. why is this newly possible now? 3. how will we reach and earn the trust of the first customers? 4. does verified customer value exceed the full cost of delivery? ai reduces the cost of execution. it does not remove the need for a customer, distribution, trust, or economics. --- # contact denys holovatyi founder & ceo osnova gmbh [linkedin](https://www.linkedin.com/in/denysholovatyi/) [email protected] --- --end--